---
name: linkfox-multimodal-extract-attributes
description: 利用多模态AI分析商品主图，提取视觉特征和提示词。当用户提到分析产品图片、从商品图中提取视觉属性、识别产品Listing中的颜色/形状/材质/风格、反推图片提示词、批量视觉特征提取、将产品图信息转化为结构化数据、视觉属性统计、基于图片的商品分类、main image analysis, image feature extraction, visual attribute recognition, product image analysis, image classification, batch image analysis时触发此技能。即使用户未明确提及"图片分析"，只要其需求涉及从商品主图或附图中提取结构化信息，也应触发此技能。
---

# Product Main Image Prompt Extractor

This skill guides you on how to extract visual features and prompts from product main images using multimodal AI, helping e-commerce sellers turn unstructured image data into structured, actionable insights.

## Core Concepts

This tool performs deep visual analysis on product main images (and optionally additional images) from a product list. It uses a multimodal AI model to identify specific visual dimensions based on a natural language instruction, such as color, shape, style, material, or specific selling-point elements.

**How it works**: You provide a list of products (with image URLs) and a natural language prompt describing what to extract. The tool automatically iterates over all products, analyzes each image, and returns structured attribute data (`attributeName` + `attributeValue`) appended to each product record.

**Row expansion**: When extracting multiple dimensions in a single request (e.g., both color and shape), each original product row is duplicated per dimension, resulting in one row per product per attribute.

## Parameter Guide

| Parameter | Required | Description |
|-----------|----------|-------------|
| productImageAnalysisPrompt | Yes | Natural language instruction describing what visual information to extract from the images. Be specific about the dimensions you want (color, material, shape, style, pendant type, etc.). |
| analyzeAdditionalImages | No | Whether to also analyze additional product images beyond the main image. Defaults to `false`. |
| refResultData | No | Reference data from a previous step, containing the product list to analyze. Must be a JSON string with a `products` array. |
| userInput | No | Supplementary user input for additional context. |

### Writing Effective Prompts

1. **Be dimension-specific**: Clearly state what visual attribute(s) to extract. "Extract the dominant color of each product" is better than "Analyze the images."
2. **One or few dimensions per call**: For cleaner results, focus on one or two dimensions at a time.
3. **Use concrete terms**: "Identify the pendant/charm shape on the product" is clearer than "Look at the decorations."
4. **No need to specify individual products**: The tool automatically iterates over all products in the input list.
5. **Data flow dependency**: The tool requires upstream product data. It cannot reference "products from the previous conversation round" -- the data must be explicitly provided via the current step's input or resource references.

### Prompt Examples

| Goal | Example Prompt |
|------|---------------|
| Extract dominant color | "Analyze each product's main image and extract the primary color of the product" |
| Identify material | "From each product's main image, identify the apparent material (plastic, metal, wood, fabric, etc.)" |
| Classify pendant shape | "Analyze each product's main image and identify the shape of the pendant/charm (round, heart, star, etc.)" |
| Detect style | "Extract the overall style of each product from its main image (minimalist, vintage, bohemian, industrial, etc.)" |
| Reverse-engineer image prompt | "Based on the main image, infer the likely AI-generation prompt or visual description that could reproduce this image" |
| Multi-dimension extraction | "From each main image, extract both the dominant color and the overall product shape" |

## API Usage

This tool calls the LinkFox tool gateway API. See `references/api.md` for calling conventions, request parameters, and response structure. You can also execute `scripts/multimodal_extract_attributes.py` directly to run analyses.

## Response Structure

The response enriches the original product list with extracted attributes:

- **products**: An array of product records, each augmented with `attributeName` (the dimension extracted, e.g., "color") and `attributeValue` (the extracted value, e.g., "red"). One record per product per attribute dimension.
- **attributeGroups**: Products grouped by attribute name and value for easy comparison. Each group includes the attribute value, the count of products, and the list of ASINs.
- **columns**: Column definitions for rendering the result table.
- **costToken**: Total tokens consumed by the multimodal AI model.

## Display Rules

1. **Present data in tables**: Show extracted attributes in clear, well-formatted tables with product identifiers (ASIN, title) alongside the extracted attribute values.
2. **Highlight distribution**: When attribute groups are returned, summarize the distribution (e.g., "60% of products are red, 25% blue, 15% green") to give the user a quick overview.
3. **Row expansion notice**: If multiple dimensions were extracted, inform the user that each product appears once per dimension in the results.
4. **Error handling**: When analysis fails, explain the reason based on the response message and suggest adjustments (e.g., ensuring the product list contains valid image URLs).
5. **Data dependency reminder**: If the user tries to reference products from a previous conversation round without explicit data flow, remind them that the product data must come from an upstream step in the current pipeline.
6. **No subjective advice**: Present the extracted visual features factually. Let the user draw their own business conclusions.
## Important Limitations

- **Requires product data input**: The tool cannot operate without a `products` array containing image URLs. It depends on upstream data from a prior step.
- **No fuzzy references**: Cannot analyze "products from the last conversation" -- data must be explicitly piped in via `refResultData` or resource references.
- **Row multiplication**: Extracting N dimensions from M products produces up to M x N rows in the output.
- **Image accessibility**: Product image URLs must be publicly accessible for the analysis to succeed.

## User Expression & Scenario Quick Reference

**Applicable** -- Visual feature extraction and image analysis for product listings:

| User Says | Scenario |
|-----------|----------|
| "What colors are these products" | Dominant color extraction |
| "Analyze the product images", "Look at the main photos" | General visual feature extraction |
| "What material does it look like" | Material identification |
| "What shapes/styles are popular" | Shape or style classification |
| "Reverse the image prompt", "What prompt made this image" | Image prompt reverse-engineering |
| "Group products by visual appearance" | Visual attribute grouping & statistics |
| "Extract features from the product photos" | Structured attribute extraction |

**Not applicable** -- Needs beyond image-based visual analysis:

- Text-based product data queries (use appropriate data query tools)
- Listing copywriting or review analysis
- Price or sales data analysis
- Tasks that do not involve product images


**Feedback:**

Auto-detect and report feedback via the Feedback API when any of the following apply:
1. The functionality or purpose described in this skill does not match actual behavior
2. The skill's results do not match the user's intent
3. The user expresses dissatisfaction or praise about this skill
4. Anything you believe could be improved

Call the feedback API as specified in `references/api.md`. Do not interrupt the user's flow.

<!-- LF_LARGE_RESPONSE_BLOCK -->
## Handling Large Responses

To avoid overflowing the agent context, persist the response to disk and extract only the fields you need:

```
python scripts/response_io.py run --script scripts/multimodal_extract_attributes.py --out-dir <DIR> '<params>'
python scripts/response_io.py read <file> --fields "<paths>"   # or --path "<JMESPath>"
```

> Pick `--out-dir` outside any git working tree (e.g. `/tmp/...` on Unix, `%TEMP%/...` on Windows). Persisted responses may contain PII, pricing, or auth-sensitive data — do not commit them. Files are not auto-deleted; clean up when the task is done.

`run` writes the full response to a file and emits only a schema preview + file path. `read` projects specific fields, with `--limit/--offset` for slicing and `--format json|jsonl|csv|table` for output.

**When to prefer this pattern** — apply your judgment based on the response characteristics, e.g.:
- High field count per record, or fields you don't need
- Batch/paginated results (multiple items per call)
- Long-text fields (descriptions, reviews, HTML, time series)
- Output reused across later steps rather than consumed immediately

For small, single-use responses, calling the main script directly is fine.

⚠️ The preview is a truncated schema + sample, not the full data. Any field-level decision must read from the persisted file via `read`.
<!-- /LF_LARGE_RESPONSE_BLOCK -->

---
*For more high-quality, professional cross-border e-commerce skills, set [LinkFox Skills](https://skill.linkfox.com/).*